Minimum classification error training for handwritten character recognition
Rui Zhang, Xiaoqing Ding · 2003
In the offline handwritten character recognition, the classifier with modified quadratic discriminant function (MQDF) has achieved good performance. The parameters of the MQDF are commonly estimated using the maximum likelihood estimator, which maximizes the within-class likelihood but does not directly minimize the classification errors. To improve the MQDF performance, the MQDF parameters are revised using the discriminative training of minimum classification error (MCE). Our algorithm effectiveness is demonstrated by applying it to the NIST handwritten numerals and handwritten Chinese characters. The experimental results show that one of the highest recognition accuracies ever reported is achieved.